This is the text extract for Letter to British Medical Journal: Is Celebrex's relative GI safety overstated?, browse documents here.

---

14 October 2002 The Editor British Medical Journal Dear Editor Celebrex’s relative GI safety is overstated? The editorial by Roger Jones1 makes important points about the limitations of the meta-analysis by Jon Deeks and colleagues2 for celecoxib. However, we also note that the Deeks meta -analysis does not account for the 12-15-month data for the CLASS study compiled by the FDA3 4 and cited by Peter Juni and colleagues’ critical editorial. 5 We have abstracted the 12-15 month CLASS data (www.fda.gov/ohrms/dockets/ac/01/briefing/3677b1_03_med.pdf tables 6, 13, 25, 26, 29, 30, www.fda.gov/ohrms/dockets/ac/01/briefing/3677b1_04_stats.doc tables 2, 5, 7) and applied them to the Deeks analysis, and find that these give a different picture. The FDA’s analysis of CLASS is more complete than that published in JAMA6, which the Juni editorial criticised for not accounting for the 12-15 month data. We believe Deeks and colleagues offer an unconvincing explanation for limiting the analysis to the six- month follow up for CLASS, insufficient to justify the post hoc changes in design, outcomes and analysis. CLASS’s 12-15 month data amount to 7878 person-years of follow-up, compared with 1252 person-years from the other RCTs measuring withdrawals because of gastrointestinal (GI) events. We also believe these data should have been included in figure 2 of the Deeks paper 'Celecoxib vs NSAID Any GI adverse effects', materially affecting those results, as shown below. Looking at withdrawals because of both serious upper GI events and endoscopic ulcers, the 12-15 month FDA data for CLASS showed no statistically significant reduction in risk (relative risk (RR) 0.73 (95% CI 0.50 to 1.05)), distinct from the 39% RRR for CLASS’s 6-month data suggested in Deeks figure 4. Likewise, for withdrawals due to serious GI events only (not endoscopic ulcers), the 12-15 month data meant the incidence of serious events (n= 20 / 3897) was nearly that of the other NSAIDS (n=24 / 3981),7 not the 11 vs. 20 effect described by Deeks for the 6-month data. 2,5 Combining the 12-month CLASS GI withdrawal data with the seven RCTs in Deeks fig 2 'Celecoxib vs NSAID Any GI adverse effects' gave an overall RRR of 41% (variance-weighted RR 0.59 (0.48 - 0.74), fixed effects method). Adjusting for the longer exposure experienced in CLASS (12-15 months rather than 12 weeks in the seven other RCTs) decreased the overall RRR to 32% (exposure/variance-weighted RR 0.68 (0.50 – 0.92)) (see Figure 1 below) - somewhat less than the 46% reported in Deeks fig 2.

---

Figure 1. Withdrawals because of adverse GI effects in celecoxib vs. NSAID RCTs

Withdrawals because of adverse GI effects in celecoxib vs NSAID RCTs

event rate control treatment 0.0% 0.5% 1.0% 1.5% 2.0% 2.5% 3.0% 3.5% 4.0% 4.5%

1.6% CLASS 12/15-month 1.2% 46/3897 vs 65/3981, RRR 27%, RR 0.73 (0.50 - 1.05) = No signf change

3.2% RCTs in Deeks + CLASS 12/15-mnth 1.9% 155/7223 vs 197/6080, RRR 41%, RR 0.59 (0.48 - 0.74) = signf decrease, ARR 1.3%, NNT 76

3.9% RCTs in Deeks + CLASS 12/15mnth, exp/var-weighted 2.6% RRR 32%, RR 0.68 (0.50 - 0.92) = signf decrease, ARR 1.2%, NNT 81

These analyses suggest celecoxib still causes statistically significant reductions in GI adverse events overall, but appreciably less than that suggested for the seven other RCTs by Deeks fig 2. Furthermore, Deeks et al reported no statistically significant difference between low-dose aspirin and non-aspirin use for both endoscopic ulcers (four RCTs, 51% vs. 73% RRR, p 0.18) and for CLASS (specific outcome not stated, 19% vs 50% RRR, p 0.44). However, using the 12-15month data for CLASS suggests a different picture. Non-aspirin users had a statistically significant 42% RRR (22 / 3154 vs 39 / 3169, RR 0.58 (0.35 - 0.95)), whereas aspirin users showed no reduction in risk (24 / 743 vs 26 / 812, RR 1.02 (0.59 - 1.74)). The difference between the subgroups’ RRRs over the 12 months was statistically significant (p 0.03). Taken in entirety (combining both endoscopic ulcers with CLASS’s GI withdrawals+ulcers)1 , significant differences between subgroups persist. When the 12-15-month CLASS data are included, the meta-analysis gives a non-significant 28% RRR for aspirin use (RR 0.72 (0.48 – 1.06)) compared with a 72% RRR for non-aspirin use (RR 0.28 (0.22 – 0.35)), the difference between RRRs being statistically significant (p <0.01). Extending the analysis to adjust for the greater exposure conferred by CLASS gave a 4% RRR for aspirin users (exposure/varianceweighted RR 0.96 (0.63 - 1.46), versus 52% for non-aspirin use (e/v-w RR 0.48 (0.33 – 0.70)), p <0.01 (see Figures 2 and 2A below):

1

Numbers of CLASS withdrawals were comp aratively low when compared with more sensitive GDUs found on 12week mandatory endoscopy in the four other RCTs. Numbers of ulcers detected by routine endoscopy at 12 weeks reported in Deeks et al figure 5 (25% control incidence) were considerably higher than numbers of withdrawals because of adverse GI effects for corresponding RCTs reported in Deeks figure 2 (6%) and in CLASS (1.6%). Hence combining the two sets of data understates ulcer burden occurring in CLASS.

---

Figure 2. Aspirin use in celecoxib vs. NSAID RCTs, GDU + withdrawals from GI events: RCTs in Deeks + CLASS 12/15-month, exposure/variance -weighted

Aspirin use in celecoxib vs NSAID RCTs, GDU + withdrawals from GI events: RCTs in Deeks + CLASS 12/15-mnth, exposure/variance-weighted

control

event rate

treatment

0%

2%

4%

6%

8%

10%

12%

10.3%

aspirin use

9.9%

42/893 vs 63/952, RRR 4%, RR 0.96 (0.63 - 1.46) = No signf change Difference between aspirin and non-aspirin RRRs p <0.01

10.4%

non-aspirin use

5.0%

90/4237 vs 273/4108, RRR 52%, RR 0.48 (0.33 - 0.70) = signf decrease, ARR 5.4%, NNT 18

Figure 2A. Aspirin use in celecoxib vs NSAID RCTs, GDU + withdrawals from GI events

Aspirin use in celecoxib vs NSAID RCTs, GDU + withdrawals from GI events

control treatment 0% 2% 4% 3.2% 3.3% 1.2% 0.7% event rate 6% 8% 10% 12%

CLASS 12/15-month, aspirin

24/743 vs 26/812, RRR -2%, RR 1.02 (0.59 - 1.74) = No signf change Difference p = 0.03

CLASS 12/15-month, non-aspirin

22/3154 vs 39/3169, RRR 42%, RR 0.58 (0.35 - 0.95) = signf decrease, ARR 0.5%, NNT 192

RCTs in Deeks + CLASS 12/15mnth, aspirin RCTs in Deeks + CLASS 12/15mnth, non-aspirin

4.7%

6.6% 42/893 vs 63/952, RRR 28%, RR 0.72 (0.48 - 1.06) = No signf change 6.6% Difference p <0.01

1.8%

90/4237 vs 273/4108, RRR 72%, RR 0.28 (0.22 - 0.35) = signf decrease, ARR 4.8%, NNT 21

RRR 4%, RR 0.96 (0.63 - 1.46) = No signf change RCTs in Deeks + CLASS 12/15mnth, aspirin, exp/var-weighted RCTs in Deeks + CLASS 12/15mnth, non-aspirin, exp/var-weighted 10.3% 9.9% Difference p <0.01 10.4% 5.0% RRR 52%, RR 0.48 (0.33 - 0.70) = signf decrease, ARR 5.4%, NNT 18

---

Hence we disagree with the implication that the benefits of celecoxib extend equally to aspirin users, and agree with NICE’s current precautionary recommendation to withhold celecoxib from aspirin users. Summary results can be seen in the Table below. Excel spreadsheet calculations and tables detailing the above findings are available on PHARMAC’s website at www.pharmac.govt.nz publications page. Methods for calculating person-year weighted incidence rates, weighted rate ratios and relative risk reductions are described in the Appendix below. We note there appears to be significant funnel plot asymmetry for the seven RCTs reported for GI withdrawals (slope 0.90, intercept 5.4, R2 0.45), with minor improvement when the CLASS results are included (see figures 3 and 4 below): Figure 3. Funnel plot asymmetry, celecoxib RCTs (withdrawals because of adverse GI effects)

Funnel plot asymmetry, withdrawals because of adverse GI effects in celecoxib vs NSAID RCTs (degree of funnel plot asymmetry measured by the intercept from regression of standard normal deviates against precision)*

*Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-analysis detected by a simple, graphical test. BMJ. 1997;315:629-34.

25.0

regression equation: SND= a+ b.(precision)

y = 0.8995x + 5.4734 R2 = 0.4561

study precision(inverse of std error)

20.0

y = 1.0314x + 4.8689 2 R = 0.5445 y = 0.8278x + 6.0702 R 2 = 0.7804

15.0

10.0

excluding CLASS including CLASS 6-month censored including CLASS 12-month total diagonal regression line, excl CLASS regression line, incl CLASS 12-month total

diagonal = funnel plot symmetry, with ideal regression line moving through 0

5.0

regression line, incl CLASS 6-month uncensored

0.0 0.0 2.0 4.0 6.0 8.0 10.0 12.0 14.0 16.0 18.0 standard normal deviate of OR (odds ratio divided by its standard error)

---

Figure 4. Funnel plot, celecoxib RCTs (withdrawals because of adverse GI effects)

Funnel plot, Withdrawals because of adverse GI effects in celecoxib vs NSAID RCTs (plot of effect estimates against sample size)

25.0 study precision (proxies sample size, as inverse of standard error) CLASS data excluding CLASS including CLASS 6-month censored 20.0 including CLASS 12-month total

15.0

10.0

5.0

favours treatment

favours controls

-1.20 -1.00 -0.80 -0.60 -0.40 -0.20 0.00 0.20 0.40 0.60 estimate of effect (ln odds ratio, where 0 = no net effect)

Such asymmetry raises the question of possible selection bias (e.g. publication bias), although might be equally explained by alternatives such as poor methodological quality of smaller studies, true heterogeneity, size of effect differs according to study size, artefact, and chance.8 This all said, results from the Success-I trial might again influence overall results.9 But future analyses must take account of the full CLASS data. In the meantime the results presented for celecoxib suggesting favourable GI safety need careful scrutiny. Finally, we note too the relatively high NNTs to prevent GI adverse events seen with celecoxib when compared with older NSAIDs (see figures 1 and 2 above), let alone negligible improvement in musculo-skeletal symptoms. Also, Cox-2 inhibitors are expensive relative to older NSAIDS in the New Zealand setting. A preliminary pharmacoeconomic analysis gives a figure of over NZ$500,000/QALY, even when using the 6-month CLASS data.

Scott Metcalfe Public Health Physician Wayne McNee Chief Executive Pharmaceutical Management Agency (PHARMAC) Level 1 Old Bank Chambers 98 Customhouse Quay PO Box 10 254 Wellington 6001 NEW ZEALAND

Sean Dougherty Analyst

---

Table. Summary results, recalculation of Deeks et al (BMJ 2002) meta-analysis of celcoxib RCTs, adverse GI effects

duration (years) total pts patients person-years measured overall variance-weighted effects (fixed effects method) (no adjustment for study duration/quality) RR

(95% CLs)

ARR info (baseline NNT event rates)

aspirin vs non-aspirin use

RRR X2 p-value ASA/nonA heteroge SA neity

RRR

signf

Withdrawals because of adverse GI effects in Celecoxib vs NSAID RCTs

RCTs in Deeks et al 2002 fig 2 'Celecoxib vs NSAID Any GI adverse effects' 7 RCTs in Deeks, which excluded CLASS* 0.23 5,425 1,252

* same as Deeks et al 2002 fig 2 'Celecoxib vs NSAID Any GI adverse effects'

0.54

(0.41,0.70)

46%

-ve

6.3%

34

CLASS 6- and 12-month results for clinically-significant upper gastrointestinal events + gastroduodenal ulcer (CSUGIE+GDU)** 6-mnth, uncensored*** 0.50 7,878 3,939 0.63 (0.41,0.98) 37% -ve 12-mnth, uncensored 1.00 7,878 0.71 (0.49,1.04) 29% 6-mnth, total (uncensored + censored) 0.65 (0.42,0.99) 35% -ve 12-mnth, total (uncensored + censored) 0.73 (0.50,1.05) 27%

**source: FDA NDA20-998 Witter J. Medical officer review. http://www.fda.gov/ohrms/dockets/ac/01/briefing/3677b1_03_med.pdf. Tables 25,26,29 *** same as Deeks et al 2002 fig 4 'Serious upper gastrointestinal events + ulcers'

1.2% 1.6% 1.3% 1.6%

220 222 220 223

combined RCTs in Deeks et al plus CLASS for [withdrawals due to any adverse GI effects] RCTs in Deeks, + CLASS 6-mnth, uncensored*** 0.39 13,303 5,191 0.56 RCTs in Deeks, + CLASS 12-mnth, total 0.59

(0.44,0.70) (0.48,0.74)

44% 41%

-ve -ve

3.7% 3.5%

61 70

Gastrointestinal impact of aspirin use in Celecoxib vs NSAID RCTs

GDU detected by routine endoscopy at 12 weeks, RCTs in Deeks et al 2002 fig 5 (excludes CLASS) non-aspirin use 0.23 2,022 467 0.27 (0.21,0.34) aspirin use 0.23 290 67 0.52 (0.30,0.87) 73% 48% -ve -ve 25.7% 25.0% 5 8

0.66 3.69 0.00

CLASS 6- and 12-month withdrawals because of clinically-significant upper gastrointestinal events + gastroduodenal ulcer (CSUGIE+GDU)* non-aspirin, 6-mnth, uncensored 0.50 6,323 3,162 0.52 (0.29,0.91) 48% -ve 1.0% aspirin, 6-mnth, uncensored 0.50 1,555 778 0.91 (0.45,1.81) 9% 2.1% non-aspirin, 12-mnth, total 1.00 6,323 6,323 0.58 (0.35,0.95) 42% -ve 1.2% aspirin, 12-mnth, total 1.00 1,555 1,555 1.02 (0.59,1.74) -2% 3.2%

*source: FDA NDA20-998 Witter J. Medical officer review. http://www.fda.gov/ohrms/dockets/ac/01/briefing/3677b1_03_med.pdf. Tables 26,30

206 517 193 -1930

0.19 -0.04

1.03 0.09 1.55 0.03

combined [RCTs in Deeks et al 12-week endoscopic GDUs] plus [CLASS CSUGIE+GDU] non-aspirin use, [CLASS 6-mnth, uncensored] 0.43 8,345 3,628 aspirin use, [CLASS 6-mnth, uncensored] 0.46 1,845 844 non-aspirin use, [CLASS 12-mnth, total] 0.81 8,345 6,790 aspirin use, [CLASS 12-mnth, total] 0.88 1,845 1,622

0.27 0.63 0.28 0.72

(0.21,0.34) (0.40,0.97) (0.22,0.35) (0.48,1.06)

73% 37% 72% 28%

-ve -ve -ve

10.0% 9.0% 8.3% 8.3%

14 30 17 43

0.51 0.39

13.90 0.00 16.85 0.00

---

Appendix: Calculating person-year weighted incidence rates, weighted rate ratios and relative risk reductions

Person-year weighted incidence rates, weighted rate ratios (relative risks) (aRR) and weighted relative risk reductions (aRRR) can be calculated as follows:

nt Nt nc Nc crude response rate for treated patients crude response rate for control patients crude rate ratio (relative risk, RR) crude odds ratio (OR) = no. of patients in treated group responding, = no. patients in treated group, = no. of patients in control group responding, = no. patients in control group = ∑nt/∑Nt = ∑nc/∑Nc = (∑nt/∑Nt) / (∑nc/∑Nc) = (∑nt/(∑Nt -∑nt)) / (∑nc/(∑Nc-∑nc))

Exposure-adjusted baseline event rates (aEc) can be derived by weighting control arms according to risk exposure (t.Nc), where t = study duration, Nc = no. patients in control group, and t.N = risk exposure, in person-year equivalents Pooled (adjusted) odds ratios for all studies (aOR) are weighted according to the variance of individual RCTs’ odds ratios, with associated confidence limits (fixed effects, Peto one-step method) Adjusted rate ratios (aRR) are derived from adjusted baseline event rates and pooled odds ratios, with associated confidence limits, according to the formula10 : aRR = 1 - (1-aEc).(1-aOR) 1 - [aEc.(1-aOR)] where: aRR = adjusted rate ratio (ie adjusted relative risk) aEc = adjusted baseline event rate (ie control incidence rate, weighted according to (t.N) aOR = pooled (adjusted) odds ratio (weighted according to variance) Adjusted relative risk reductions (aRRR) are derived from adjusted rate ratios, where aRRR = 1 – aRR, according to the formula: aRRR = (1-aEc).(1-aOR) 1 - [aEc.(1-aOR)] where: aRRR = adjusted relative risk reduction aRR = adjusted rate ratio aEc = adjusted baseline event rate aOR = pooled (adjusted) odds ratio If adjusted baseline event rates are considered relevant to the New Zealand population, adjusted absolute risk reductions (aARR) are derived from adjusted baseline event rates and adjusted rate ratios, according to the formula: aARR = aEc.aRRR where: aARR = adjusted absolute risk reduction aEc = adjusted baseline event rate aRRR = adjusted relative risk reduction Similarly, if adjusted baseline event rates are considered relevant to the New Zealand population, adjusted treatment event rates (aEt) are derived from adjusted baseline event rates and adjusted rate ratios, according to the formula: aEt = aEc. aRR where: aEt = adjusted treatment event rate (treated incidence rate) aEc = adjusted baseline event rate aRR = adjusted rate ratio (relative risk, derived from pooled odds ratio)

---

[NB Odds ratios derive from relative risks according to the formula: OR = RR.(1-Ec) 1 – (RR. Ec ) = where: OR RR Ec Et = odds ratio = rate ratio (ie relative risk) = baseline (control) event rate = treatment event rate Et .((1/Ec)-1) 1- Et

]

⇒

To account for the quality of contributing RCTs, each RCT can be weighted according to a quality score, for example PHARMAC ’s following modification of the Jadad criteria 11 (score 0-5): Criterion (modified) Randomisation Concealed allocation Blinding of receipt Source of bias (Cochrane Handbook taxonomy) Selection bias / confounding, i.e. systematic differences in comparison groups Selection bias / confounding Scoring system Adequate =1, Inadequate/nil = 0 Adequate =1, Unclear/not described = 0 Inadequate/nil = 0 Adequate, described =0.5, Unclear/not described = 0.25 Inadequate/nil = 0 Adequate, described =0.5, Unclear/not described = 0.25 Inadequate/nil = 0 Participants adequately accounted for =1, Unclear/not described = 0 Inadequate/nil = 0 0 (presumably incorporated into Blinding of provision)

Blinding provision Follow-up

of

Performance bias, i.e. systematic differences in care provided apart from the intervention being evaluated; recipients Performance bias; providers

Attrition bias, i.e. systematic differences in withdrawals from the trial, affecting outcome measurement of Detection bias, i.e. systematic differences in outcome assessment; assessors

Blinding assessment

Combining these quality-based weights with the above [variance-based weights for odds ratios] and the [exposure-based (t.N) weights for adjusted baseline incidence rates] gives quality/variance weights and quality/exposure/variance weights. These quality-containing weights can be used to then calculate qualityweighted pooled odds ratios and quality-weighted adjusted baseline incidence rates, using the above formulae, and thus quality-weighted adjusted rate ratios, etc.

---

References

1

Jones R. Efficacy and safety of COX 2 inhibitors. BMJ 2002;325:607-608. http://bmj.com/cgi/content/full/325/7365/607

2

Deeks JJ, Smith LA, Bradley MD. Efficacy, tolerability, and upper gastrointestinal safety of celecoxib for treatment of osteoarthritis and rheumatoid arthritis: systematic review of randomised controlled trials. BMJ 2002; 325: 619-623. http://bmj.com/cgi/content/full/325/7365/619

3

Witter J. Medical officer review. www.fda.gov/ohrms/dockets/ac/01/briefing/3677b1_03_med.pdf

4

Lu HL. Statistical reviewer briefing document for the advisory committee. www.fda.gov/ohrms/dockets/ac/01/briefing/3677b1_04_stats.doc

5

Juni P, Rutjes AWS, Dieppe PA. Are selective COX-2 inhibitors superior to traditional non-steroidal antiinflammatory drugs? BMJ 2002; 324: 1287-1288. http://bmj.com/cgi/content/full/324/7349/1287

6

Silverstein FE, Faich G, Goldstein JL, Simon LS, Pincus T, Whelton A, et al. Gastrointestinal toxicity with celecoxib vs nonsteroidal anti-inflammatory drugs for osteoarthritis and rheumatoid arthritis: the CLASS study: A randomised controlled trial. JAMA 2000; 284: 1247-1255.

7

Bassett K, Wright JM, Puil L, Perry TL Jr, Heran B, Cole C. Cyclooxygenase-2 inhibitor update: Journal articles fail to tell the full story. Can Fam Physician 2002;48:1455-1460.

8

Clarke M, Oxman AD, editors. Formulating the problem. Cochrane Reviewers' Handbook 4.1.5 [updated April 2002]; Section 8.11. In: The Cochrane Library, Issue 2, 2002. Oxford: Update Software. Updated quarterly. http://www.cochrane.dk/cochrane/handbook/hbook811_Special_topics.htm

9

Singh G, Goldstein J, Fort J, Bello A, Boots S. Success-I in osteoarthritis (OA) trial: celecoxib has similar efficacy to the conventional NSAIDS [abstract]. J Rheumatol 2001; 28(suppl 6.3): 6.

10

algebraic transformation by PHARMAC of formulae in Sackett D, Straus S, Richardson WS, Rosenberg W, Haynes B. Evidence-based medicine: how to practice and teach EBM, 2nd edition. Oxford: Churchill Livingstone, 2000. p136 Table 5.1 Formulae to convert odds ratios (ORs) and relative risks (RRs) to NNTs.

11

Jadad AR, Moore RA, Carroll D, Jenkinson C, Reynolds DJ, Gavaghan DJ, McQuay HJ. Assessing the quality of reports of randomized clinical trials: is blinding necessary? Control Clin Trials. 1996;17:1-12.